Playoff Odds & Win Curve
20,000 simulated seasons from each club’s controlled roster, with measured uncertainty.
What it predicts
Probability of reaching the postseason (and of winning the division).
Inputs
- Each club’s current 40-man minus its pending free agents, run through the depth chart (who plays where, playing time by role)
- The balanced schedule and the 12-team playoff format
Method
- Team strength = replacement level + WAR from the depth chart, each player at his own role’s playing time with no team cap (the Winter Plan and Roster Builder hand out one budgeted season and measure changes from this number; the budget backtested no better)
- Talent uncertainty measured on 150 team-seasons (2022–26, each club’s opening-day roster, each season forecast blind), after removing game-to-game luck
- Log5 game probabilities, every game simulated; common random numbers: every point on a club’s win curve re-plays the same 20,000 seasons, so the curve is smooth and its slope isn’t simulation noise; a probit fit to those points, anchored to the club’s own simulated odds
- Unsigned free agents (their projected WAR) go to the other 29 clubs by expected spending: half payroll-predicted (tested on the 2025–26 winter), half an even split; any remaining gap to 2,430 league wins is shared by all 30
How well it works
Computed by the pipeline from the data on this site
- Backtest: every club’s opening-day roster, 2022–26, each season forecast blind
- r = 0.652, RMSE 9.5 wins (150 team-seasons) Best 13 + 13 (the old method): r = 0.644, RMSE 9.53. Marcel playing time: r = 0.655, RMSE 9.41. Last season’s BaseRuns record, no roster at all: RMSE 9.92.
- Playoff odds, backtest (150 team-seasons, 2022–26): Brier score, lower is better
- 0.198 vs. 0.240 for a flat 40% for every club 10% predicted → 10% made it (40) · 31% predicted → 45% made it (33) · 50% predicted → 45% made it (42) · 66% predicted → 50% made it (26) · 86% predicted → 100% made it (9). Mildly in-sample: the talent spread was measured on the same seasons.
- Uncertainty used in every simulation
- ±7.0 wins of talent + ±6.4 wins of game-to-game luck
What it can't see
- Free agency hasn’t happened; payroll predicts spending only loosely, so real signings will be lumpier.
- No team playing-time cap: a deep 40-man can claim more than one season of plate appearances and innings.
- No roster-need term: a club’s free-agent share ignores its open playing time, and a surplus at one position isn’t discounted; testing one needs October rosters from past winters, which Keystone hasn’t collected.
- A two-way player counts as a pitcher only, so his bat is left out of his club’s projection.
- Mid-season trades and injuries beyond the measured uncertainty.
- Dollars per win need private revenue data.
How we'd measure impact
- Calibration of preseason odds vs. outcomes over several seasons (the 2022–26 backtest: Brier score against a flat guess, on the card)
- Team-win error vs. Marcel playing time and last season’s BaseRuns record